Pith. sign in

REVIEW 6 major objections 5 minor 34 references

SemRAG: Semantic Knowledge-Augmented RAG for Improved Question-Answering

T0 review · 6 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read SemRAG claims that semantic chunking plus knowledge-graph community retrieval lifts RAG answer relevance and correctness by up to 25% over naive RAG, without fine-tuning.

desk verdict A legitimate GraphRAG-plus-semantic-chunking pipeline whose central empirical claims are contradicted by its own reported tables and are not reproducible as stated. read the letter →

arxiv 2507.21110 v1 pith:TPXJVSQY submitted 2025-07-10 cs.CL cs.AIcs.IR

Kezhen Zhong , Basem Suleiman , Abdelkarim Erradi , Shijing Chen This is my paper
classification cs.CLcs.AIcs.IR
keywords semanticchunkingretrieval-augmentedgenerationknowledgegraphcommunityretrievalcosinesimilaritybuffersizeanswerrelevancemulti-hopquestionanswering
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes SemRAG, a retrieval-augmented generation pipeline that replaces fixed-size chunking with semantic chunking: adjacent sentences are grouped into a chunk when the cosine distance between their embeddings falls below a threshold, with a buffer parameter controlling how many neighbors are merged. Those chunks are then used to build a hierarchical knowledge graph whose communities are retrieved locally or globally to answer a query. The paper's claim is that this combination improves answer relevance and correctness, by up to 25% in certain configurations on the MultiHopRAG and Wikipedia benchmarks, while avoiding the cost of fine-tuning. The practical payoff would be a cheaper, scalable route to domain-specific accuracy in LLM question answering.

What carries the argument

The load-bearing mechanism is the semantic chunking algorithm: a chunk $g$ is the set of neighboring sentence groups $\{c_i \mid 1 - \frac{d(c_i)\cdot d(c_{i+k})}{\|d(c_i)\|_2\,\|d(c_{i+k})\|_2} < \tau \text{ for all } k \in [0,n]\}$, so adjacent sentences stay together while their embedding cosine distance stays below a threshold. A buffer size $b$ merges context around each sentence before embedding, and chunks over the token limit are split into overlapping sub-chunks. These chunks feed knowledge-graph construction with community detection, and retrieval selects either local entities and chunks by similarity or global community summaries ranked against the query. The paper's claim is that this combination preserves both sentence-level coherence and relational structure, which is what improves answer relevance and correctness.

What would settle it

Run the same knowledge-graph retrieval with fixed-size chunks instead of semantic chunks: if the reported answer-relevance gain disappears, semantic chunking is the active ingredient; if it persists, the gain comes from graph retrieval. Separately, comparing the chunk boundaries produced by Algorithm 1 against gold topic-segmentation boundaries on a corpus would show whether cosine-distance splits track true semantic structure.

Watch

Extended reading notes

Core claim

On the paper's own terms, SemRAG establishes that indexing documents by semantically coherent chunks rather than fixed-size blocks, and then retrieving from knowledge-graph communities built on those chunks, yields more relevant and more correct answers than naive RAG for the tested models and datasets. The authors report that semantic chunking at buffer size 0 improves answer relevancy for all tested models relative to naive RAG on MultiHopRAG, and that the best buffer size is corpus-dependent, with MultiHop peaking at buffer 5 and Wikipedia at buffer 12 for correctness. They also report an 11% to 12% answer-relevancy improvement over conventional RAG in their conclusion. The contribution is framed as an efficient, fine-tuning-free integration of domain knowledge into LLMs.

Load-bearing premise

The pipeline assumes cosine similarity between embeddings of adjacent sentences is a reliable signal for where semantically coherent chunks should end, and this premise is never checked against ground-truth chunk boundaries.

Editorial extensions

If this is right

  • Semantic chunking with buffer size 0 raises answer relevancy for all tested models on MultiHopRAG compared with naive RAG.
  • Optimal buffer size is corpus-specific: MultiHop peaks at buffer 5, Wikipedia at buffer 12, and larger buffers eventually hurt correctness or relevance.
  • Larger buffers increase knowledge-graph size and construction time roughly linearly, so chunking must balance context richness against computational cost.
  • The full pipeline needs no fine-tuning, so the claimed gains are available to small local models in domain-specific settings.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper never validates that cosine similarity between sentence embeddings marks true semantic boundaries, so the central gain may depend on the embedding model; a ground-truth chunk-boundary metric would test this directly.
  • Because only one embedding model is used throughout, it is an open question whether the optimal buffer sizes transfer across embedders; the framework implies buffer tuning per corpus, not per model.
  • Some small-buffer configurations collapse to 'insufficient information' responses, suggesting retrieval sufficiency rather than chunk quality may be the binding constraint in those regimes.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

6 major / 5 minor

Summary. The paper proposes SemRAG, a retrieval-augmented generation pipeline that combines semantic chunking (grouping sentences by cosine-similarity of their embeddings) with knowledge-graph construction and community-based retrieval. It is evaluated on MultiHopRAG and RAG Mini Wikipedia using Mistral, Llama3, and Gemma2, with RAGAS answer-correctness, answer-similarity, and answer-relevancy metrics. The paper claims improvements over naive RAG of up to 25% and reports a buffer-size optimization study. The experimental evidence in the manuscript does not consistently support these claims, and the reported tables contain internal contradictions.

Significance. A well-validated lightweight semantic-chunking-plus-knowledge-graph RAG method would be practically useful, and the proposed design is plausible. The paper, however, provides no code, no full configuration details, and its own tables conflict on central measurements. The claimed improvements are not consistent across models and metrics, the buffer-size 'optimum' is a post-hoc argmax without uncertainty quantification, and the semantic-chunking threshold is never validated against a ground-truth chunk metric. These issues place the entire empirical contribution in question.

major comments (6)
  1. [§4.3, Table 5] The stated MultiHop optimum of buffer size 5 is contradicted by Table 5: Answer Correctness at Buffer 0 is 0.340 versus 0.320 at Buffer 5, and Answer Relevancy differs by only 0.004 (0.575 vs 0.571), while Table 3 reports standard errors of roughly 0.012–0.017. The claimed peak is therefore within the noise floor, and the buffer-size conclusion in Section 4.3 and Figure 9 is not supported by the paper's own data.
  2. [Table 3, Abstract] The headline 'up to 25%' improvement is not traceable to a consistent comparison. Semantic Chunk(0) reduces Mistral Answer Correctness relative to Naive (0.284 vs 0.305), and Semantic Chunk(5) reduces Gemma2 Answer Relevancy relative to Naive (0.352 vs 0.505). The paper does not report significance tests or a consistent model/metric selection rule, so the gain appears to be a selected cell rather than a systematic effect.
  3. [Tables 1 and 2] The same MultiHop Buffer-0 configuration is reported with conflicting graph statistics: Table 1 lists Nodes 861, Edges 338, Time 3613s, while Table 2's Mistral row lists Nodes 170, Edges 65, Time 8166s (the Table 1 values match the Llama3 row in Table 2). This makes the experimental measurements unreproducible and raises concerns about the accuracy of the reported runtime and graph-complexity analyses.
  4. [§3.2.2, Eq. (1), Algorithm 1] The chunking threshold is a free parameter with inconsistent notation (τ vs θ) and is never validated against a ground-truth chunk metric; the paper itself acknowledges in Future Work that such a metric is still needed. Without this validation, the reported retrieval improvements cannot be attributed to semantic coherence rather than to threshold selection.
  5. [§4.3, Figures 7–9] The buffer-size optimization is circular as presented: buffer values are scanned on each dataset, the best-scoring value is selected, and that argmax is then reported as the 'optimal' buffer size. There is no held-out validation or out-of-sample prediction, so the conclusion is a fitted maximum rather than a demonstrated property of the method.
  6. [Entire manuscript] No code or full hyperparameter configuration is provided anywhere in the manuscript, so even the internally consistent parts of the evaluation cannot be reproduced by a third party. Given that the central claim rests entirely on these empirical results, the absence of a reproducible experimental protocol is a load-bearing problem.
minor comments (5)
  1. [Figure 9] Figure 9 is labeled 'Mixtral' while Section 4.3 and Figure 7 discuss Llama 3; the model used for the optimal-buffer experiments should be stated consistently.
  2. [§3.2.2, Eq. (1), Algorithm 1] Equation (1) defines a chunk g using a condition over all k in [0,n] and threshold τ, but Algorithm 1 uses a threshold θ on adjacent pairwise cosine distances; the relationship between these definitions is not explained.
  3. [Abstract] The abstract contains grammatical errors, e.g., 'SemRAG has significantly enhances,' and similar typos appear throughout (e.g., 'tradiational,' 'Naïve' vs 'Naive').
  4. [Section 4, references] Citation callouts such as '2[24]' and a bare '3' in the dataset descriptions are formatting errors; the citation syntax should be corrected throughout.
  5. [Section 5] The conclusion states an '11% to 12% improvement in answer relevancy,' which is inconsistent with the abstract's 'up to 25%' claim; the authors should reconcile these numbers with exact table cells.

Circularity Check

1 steps flagged · score 6.0 of 10

The dataset-specific buffer-size 'optimum' is an argmax over the scanned grid restated as a finding; the core RAG comparison remains an empirical benchmark rather than a derived prediction.

  1. fitted input called prediction [Section 4.3, Figure 9; Tables 4 and 5]
    "Figure 9 highlights a non-linear relationship between buffer size and RAG performance, particularly in the metrics of answer relevancy and correctness. Contrary to the assumption that increasing buffer size uniformly enhances model outputs, results reveal dataset-specific optima: the Multi-hop dataset reaches peak performance at buffer size 5, while the Wiki dataset achieves its highest scores at size 10."

    The stated optima are the maxima of the RAGAS score arrays already tabulated in Tables 4-5 for the same evaluation sets; no held-out split or independent validation is described. 'Buffer 5 is optimal for MultiHop' is therefore an argmax over the scanned grid restated as a finding, i.e., a fitted selection rather than an independent prediction. The reduction is visible in Table 5, where the text's chosen optimum (buffer 5, correctness 0.320) is not even the array maximum (buffer 0, correctness 0.340), so the claimed optimum is selected rather than derived from the reported scores.

full rationale

The central comparison between SemRAG and Naive/Fixed-Size RAG is an empirical benchmark on MultiHopRAG and Wikipedia, and that comparison is not circular: the RAGAS metrics are external to the chunking definition, and no self-citation or imported uniqueness theorem carries the argument. The partial circularity is confined to the buffer-size optimization contribution. The paper scans buffer values 0-10 (and up to 30 for Wiki), measures scores on the same test sets, and then reports the argmax as a discovered 'optimal' buffer size. Because no validation split or independent ground-truth chunk metric is provided, this conclusion reduces to a fitted maximum by construction; it is a selection criterion, not a prediction. The paper itself notes in Future Work that 'developing a ground-truth metric for evaluating chunk boundaries' is needed, which underscores the absence of an external anchor for the chunking claim. Separately, internal inconsistencies (e.g., Table 1 vs Table 2 node/edge/time values, and Table 5 failing to support the claimed buffer-5 peak) raise reproducibility concerns, but these are correctness risks rather than circularity. Overall, the headline method comparison retains independent empirical content, while one claimed result--the corpus-specific optimal buffer--is a restatement of the test-set argmax.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

No new free parameters beyond those listed; the central mechanism is empirical. The paper contributes no first-principles derivation, so all claims rest on the assumptions above and on a buffer-size scan that is fitted to each dataset.

free parameters (4)
  • Cosine distance threshold theta/tau = Not reported
    Eq. (1) and Algorithm 1 use this threshold to decide chunk boundaries; no value is given anywhere.
  • Buffer size b = 5 for MultiHop, 10-12 for Wiki depending on table
    Scanned per dataset to find the reported optimum; this is a fitted hyperparameter, not a prediction.
  • Sub-chunk token limit and overlap = 1024 tokens, overlap 128 tokens
    Chosen by design in Eq. (2) without ablation, affecting all chunk compositions.
  • Retrieval window L, top-k, thresholds tau_e and tau_d = Not reported
    Eqs. (4) and (5) define retrieval using these values, but the paper never states them.
assumptions (4)
  • domain assumption Cosine similarity between sentence embeddings captures semantic coherence for chunking.
    Section 3.2.2 and Eq. (1) rest on this; no ground-truth chunk metric is provided.
  • domain assumption Knowledge graph community summaries are more useful for answer generation than raw chunks.
    Section 3.2.3 and Eq. (3) assume this; no ablation isolates the KG contribution.
  • domain assumption RAGAS scores from ChatGPT-4o-mini are valid measures of correctness, similarity, and relevance.
    Section 4.1 relies on this metric without human validation or agreement analysis.
  • domain assumption The Microsoft GraphRAG components (entity extraction, Leiden communities, report summarization) behave as described.
    Sections 3.1 and 3.2 import ref [21] without independent verification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SemRAG: Semantic Knowledge-Augmented RAG for Improved Question-Answering." pith.science (2026). https://pith.science/paper/TPXJVSQY

@misc{pith2026250721110,
  author       = {Pith},
  title        = {Pith review of: SemRAG: Semantic Knowledge-Augmented RAG for Improved Question-Answering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TPXJVSQY}},
  note         = {Machine review of arXiv:2507.21110}
}
read the original abstract

This paper introduces SemRAG, an enhanced Retrieval Augmented Generation (RAG) framework that efficiently integrates domain-specific knowledge using semantic chunking and knowledge graphs without extensive fine-tuning. Integrating domain-specific knowledge into large language models (LLMs) is crucial for improving their performance in specialized tasks. Yet, existing adaptations are computationally expensive, prone to overfitting and limit scalability. To address these challenges, SemRAG employs a semantic chunking algorithm that segments documents based on the cosine similarity from sentence embeddings, preserving semantic coherence while reducing computational overhead. Additionally, by structuring retrieved information into knowledge graphs, SemRAG captures relationships between entities, improving retrieval accuracy and contextual understanding. Experimental results on MultiHop RAG and Wikipedia datasets demonstrate SemRAG has significantly enhances the relevance and correctness of retrieved information from the Knowledge Graph, outperforming traditional RAG methods. Furthermore, we investigate the optimization of buffer sizes for different data corpus, as optimizing buffer sizes tailored to specific datasets can further improve retrieval performance, as integration of knowledge graphs strengthens entity relationships for better contextual comprehension. The primary advantage of SemRAG is its ability to create an efficient, accurate domain-specific LLM pipeline while avoiding resource-intensive fine-tuning. This makes it a practical and scalable approach aligned with sustainability goals, offering a viable solution for AI applications in domain-specific fields.

Figures

Figures reproduced from arXiv: 2507.21110 by the authors.

Figure 1
Figure 1. SemRAG framework leveraging semantic chunking and knowledge graphs for enhanced contextual under￾standing; (a) Semantic Indexing: Segment the document into smaller chunks indexed and stored in a database. (b) Context Retrieval: Retrieve the top k community based on semantic similarity. (c) Knowledge Graph: Information is extracted to build the knowledge graph hierarchy with communities. optimal chunk size presents a… view at source ↗
Figure 2
Figure 2. SemRAG consist of two phases, semantic indexing and graph communities construction. (a) Semantic Chunking: long text has been split and segmented into meaningful chunks based on semantic relevance. (b) Graph Communities Construction: building a knowledge graph based on hierarchical community summarization. As shown in figure 2, the semantic chunking algorithm works by first splitting the input text into individual s… view at source ↗
Figure 3
Figure 3. Comparison of LLMs and Mistral Performance with Different Buffer Sizes (AS = Answer Similarity, AR = Answer Relevancy, AC = Answer Correctness) 4.2 Model Performance with SemRAG As shown in [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: RAGAS metrics comparison with Multi Hop-RAG (Naive RAG vs Semantic Chunking) Based on [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: RAGAS metrics comparison with Multi Hop-RAG (Fixed Size vs Semantic Chunking with Knowledge Graph Integration insufficient information retrieval by the RAG. As a result, these models frequently generated "Insufficient Information" outputs, particularly in MultiHopRAG t…
Figure 6
Figure 6. Figure 6: Buffer Size vs Graph Complexity • Buffer size has a direct and measurable impact on graph complexity As buffer size increases, both the number of nodes and edges in the resulting knowledge graphs scale linearly, reflecting the expanded scope of retrieved context. This …
Figure 7
Figure 7. Figure 7: Performance of Buffer Sizes (0-10) vs Naive RAG Based on RAGAS Test Metrics Llama 3 (Multi-Hop) 0 1 2 3 4 5 6 7 8 9 10 12 20 25 30 Buffer Size 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Metric Score Answer Correctness Answer Similarity Answer Relevancy Answer Correctness Naiv…
Figure 8
Figure 8. Figure 8: Performance of Buffer Sizes (0-30) vs Naive RAG Based on RAGAS Test Metrics Llama 3 (wiki) sizes for both semantic chunking and Naive RAG, suggesting minimal sensitivity to buffer variation. Answer Relevancy shows moderate fluctuations but stays near 0.7 for semantic c…
Figure 9
Figure 9. Figure 9: Optimal Chunk Comparison (Mixtral) (Multi-hop vs Wiki Dataset) benefit from enriched contextual information that supports improved reasoning and content generation. However, beyond these optimal points, performance either plateaus or declines. This is likely due to inf…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 25 canonical work pages

  1. [1]

    Large legal fictions: Profiling legal hallucinations in large language models

    Matthew Dahl, Varun Magesh, Mirac Suzgun, and Daniel E Ho. Large legal fictions: Profiling legal hallucinations in large language models. Journal of Legal Analysis, 16(1):64–93, 06 2024

    work page 2024

  2. [2]

    Retrieval-augmented generation for knowledge-intensive nlp tasks

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. Retrieval-augmented generation for knowledge-intensive nlp tasks. In Advances in Neural Information Processing Systems, volume 33, pages 9459–9474. Curran Associates, ...

    work page 2020

  3. [3]

    Retrieval-augmented generation for large language models: A survey

    Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, Meng Wang, and Haofen Wang. Retrieval-augmented generation for large language models: A survey. arXiv, 2023. arXiv:2312.10997 [cs.CL] https://arxiv.org/abs/2312.10997

    arXiv 2023

  4. [4]

    Yue et al. Yu. Rankrag: Unifying context ranking with retrieval-augmented generation in llms. In Advances in Neural Information Processing Systems, volume 37, pages 121156–121184. Curran Associates, Inc., 2024

    work page 2024

  5. [5]

    Is semantic chunking worth the computational cost? In Findings of the Association for Computational Linguistics: NAACL 2025, pages 2155–2177

    Renyi Qu, Ruixuan Tu, and Forrest Sheng Bao. Is semantic chunking worth the computational cost? In Findings of the Association for Computational Linguistics: NAACL 2025, pages 2155–2177. Association for Computational Linguistics, April 2025

    work page 2025

  6. [6]

    NLTK: The natural language toolkit

    Steven Bird and Edward Loper. NLTK: The natural language toolkit. In Proceedings of the ACL Interactive Poster and Demonstration Sessions, pages 214–217, Barcelona, Spain, July 2004. Association for Computational Linguistics

    2004

  7. [7]

    Knowledge graphs

    Aidan Hogan, Eva Blomqvist, Michael Cochez, and et al. Knowledge graphs. ACM Computing Surveys , 54(4):71:1–71:37, 2021

    work page 2021

  8. [8]

    On the effectiveness of parameter-efficient fine-tuning

    Zihao Fu, Haoran Yang, Anthony Man-Cho So, and et al. On the effectiveness of parameter-efficient fine-tuning. In Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence, 2023

    work page 2023

Show all 34 references
  1. [9]

    Qingxiu et al. Dong. A survey on in-context learning. In Proceedings of EMNLP 2024, pages 1107–1128, Miami, USA, 2024. ACL

    2024

  2. [10]

    In-context retrieval-augmented language models

    Ori Ram, Yoav Levine, Itay Dalmedigos, Dor Muhlgay, Amnon Shashua, Kevin Leyton-Brown, and Yoav Shoham. In-context retrieval-augmented language models. Trans. Assoc. Comput. Linguistics, 11:1316–1331, 2023

    2023

  3. [11]

    Blended rag: Improving rag (retriever-augmented generation) accuracy with semantic search and hybrid query-based retrievers

    Kunal Sawarkar, Abhilasha Mangal, and Shivam Raj Solanki. Blended rag: Improving rag (retriever-augmented generation) accuracy with semantic search and hybrid query-based retrievers. In 2024 IEEE 7th International Conference on Multimedia Information Processing and Retrieval (...

    2024

  4. [12]

    Nguyen Nam et al. Doan. A hybrid retrieval approach for advancing retrieval-augmented generation systems. In Proc. ICNLSP 2024, pages 397–409, Trento, 2024. ACL

    2024

  5. [13]

    DSLR: Document refinement with sentence-level re-ranking and reconstruction to enhance retrieval-augmented generation

    Taeho Hwang, Soyeong Jeong, Sukmin Cho, SeungYoon Han, and Jong Park. DSLR: Document refinement with sentence-level re-ranking and reconstruction to enhance retrieval-augmented generation. In Wenhao Yu, Weijia Shi, Michihiro Yasunaga, Meng Jiang, Chenguang Zhu, Hannaneh Hajish...

    2024

  6. [14]

    Large language models are effective text rankers with pairwise ranking prompting

    Zhen Qin, Rolf Jagerman, Kai Hui, Honglei Zhuang, Junru Wu, Le Yan, Jiaming Shen, Tianqi Liu, Jialu Liu, Donald Metzler, Xuanhui Wang, and Michael Bendersky. Large language models are effective text rankers with pairwise ranking prompting. In Kevin Duh, Helena Gomez, and Steve...

    2024

  7. [15]

    Large language models and future of information retrieval: Opportunities and challenges

    ChengXiang Zhai. Large language models and future of information retrieval: Opportunities and challenges. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’24, page 481–490, New York, NY , USA, 2024. Asso...

    2024

  8. [16]

    RAGGED: Towards informed design of retrieval augmented generation systems, 2025

    Jennifer Hsia, Afreen Shaikh, Zhiruo Wang, and Graham Neubig. RAGGED: Towards informed design of retrieval augmented generation systems, 2025. arXiv:2403.09040 [cs.CL] https://arxiv.org/abs/2403.09040

    2025 arXiv

  9. [17]

    Mix-of-granularity: Optimize the chunking granularity for retrieval-augmented generation

    Zijie Zhong, Hanwen Liu, Xiaoya Cui, Xiaofan Zhang, and Zengchang Qin. Mix-of-granularity: Optimize the chunking granularity for retrieval-augmented generation. In Owen Rambow, Leo Wanner, Marianna Apidianaki, Hend Al-Khalifa, Barbara Di Eugenio, and Steven Schockaert, editors...

    2025

  10. [18]

    Enhancing retrieval augmented generation with hierar- chical text segmentation chunking

    Hai-Toan Nguyen, Tien-Dat Nguyen, and Viet-Ha Nguyen. Enhancing retrieval augmented generation with hierar- chical text segmentation chunking. In Information and Communication Technology, pages 209–220. Springer Nature Singapore, 2025

    2025

  11. [19]

    Graph retrieval-augmented generation for large language models: A survey

    Tyler Thomas Procko and Omar Ochoa. Graph retrieval-augmented generation for large language models: A survey. In 2024 Conference on AI, Science, Engineering, and Technology (AIxSET), pages 166–169, 2024

    2024

  12. [20]

    Xiangrong et al. Zhu. Knowledge graph-guided retrieval augmented generation. In Proc. NAACL-HLT 2025 (Vol. 1: Long Papers), pages 8912–8924, Albuquerque, New Mexico, 2025

    2025

  13. [21]

    Ness, and James Larson

    David Edge, Huy Trinh, Nancy Cheng, Jeff Bradley, Albert Chao, Abhishek Mody, Stephen Truitt, David Metropolitansky, Ross O. Ness, and James Larson. From local to global: A graph rag approach to query-focused summarization, 2024. arXiv:2404.16130 [cs.CL] https://arxiv.org/abs/...

    2024 arXiv

  14. [22]

    Mariam et al. Barry. GraphRAG: Leveraging graph-based efficiency to minimize hallucinations in LLM-driven RAG for finance data. In Proc. Workshop on Generative AI and Knowledge Graphs (GenAIK), pages 54–65, Abu Dhabi, UAE, 2025. ICCL

    2025

  15. [23]

    Salton, A

    G. Salton, A. Wong, and C. S. Yang. A vector space model for automatic indexing. Commun. ACM, 18(11):613– 620, 1975

    1975

  16. [24]

    Multihop-RAG: Benchmarking retrieval-augmented generation for multi-hop queries

    Yixuan Tang and Yi Yang. Multihop-RAG: Benchmarking retrieval-augmented generation for multi-hop queries. In First Conference on Language Modeling, 2024

    2024

  17. [25]

    Jiang, Alexandre Sablayrolles, Arthur Mensch, et al

    Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, et al. Mistral 7b, 2023. arXiv:2310.06825 [cs.CL] https://arxiv.org/abs/2310.06825

    2023 arXiv

  18. [26]

    The llama 3 herd of models,

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, and et al. The llama 3 herd of models,

  19. [27]

    Gemma 2: Improving open language models at a practical size, 2024

    Gemma Team at Google DeepMind. Gemma 2: Improving open language models at a practical size, 2024. arXiv:2408.00118 [cs.CL] https://arxiv.org/abs/2408.00118

    2024 arXiv

  20. [28]

    Morris, Brandon Duderstadt, and Andriy Mulyar

    Zach Nussbaum, John X. Morris, Brandon Duderstadt, and Andriy Mulyar. Nomic embed: Training a reproducible long context text embedder, 2024. Accessed: 2024-10-18

    2024

  21. [29]

    RAGAs: Automated evaluation of retrieval augmented generation

    Shahul Es, Jithin James, Luis Espinosa Anke, and Steven Schockaert. RAGAs: Automated evaluation of retrieval augmented generation. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics: System Demonstrations, pages 150–158, 2024

    2024

  22. [30]

    Speculative RAG: Enhancing retrieval augmented generation through drafting

    Zilong Wang, Zifeng Wang, Long Le, Steven Zheng, Swaroop Mishra, Vincent Perot, Yuwei Zhang, Anush Mattapalli, Ankur Taly, Jingbo Shang, Chen-Yu Lee, and Tomas Pfister. Speculative RAG: Enhancing retrieval augmented generation through drafting. In The Thirteenth International ...

    2025

  23. [31]

    Tanyildiz

    Derya et al. Tanyildiz. Enhancing retrieval-augmented generation accuracy with dynamic chunking and optimized vector search. Orclever Proc. Res. Dev., 5(1):215–225, December 2024

    2024

  24. [32]

    Lightrag: Lightweight retrieval-augmented generation framework, 2024

    HKU Data Science Lab. Lightrag: Lightweight retrieval-augmented generation framework, 2024. arXiv:2410.05779 [cs.CL] https://arxiv.org/abs/2410.05779

    2024 arXiv

  25. [33]

    Tong Chen, Hongwei Wang, Sihao Chen, Wenhao Yu, Kaixin Ma, Xinran Zhao, Hongming Zhang, and Dong Yu. Dense X retrieval: What retrieval granularity should we use? In Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen, editors, Proceedings of the 2024 Conference on Empirical Meth...

    2024

  26. [2024]

    arXiv:2407.21783 [cs.CL] https://arxiv.org/abs/2407.21783

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.